PostgreSQL außer Kontrolle geratene IO

[JanJoost]

Hallo zusammen,

ich habe ein Problem, bei dem ich eure Hilfe benötige. Ich habe einen Server mit Discourse installiert, der auf dem Docker-basierten Installer läuft. Jeden Tag gegen Mittag wird der Server aus mir noch unbekannten Gründen langsam, und ich sehe, dass die I/O-Last so stark ansteigt, dass das Forum anfängt, Fehler 500 anzuzeigen.

iotop meldet mir stets, dass es sich um ein PostgreSQL-UPDATE handelt, das die gesamte I/O-Last aufbraucht.

Als es heute wieder passiert ist, habe ich mit diesem Befehl eine Liste aller laufenden SQL-Abfragen erstellt:
sudo -u postgres psql discourse -o /tmp/RunningQueries -c ‘SELECT datname,pid,state,query FROM pg_stat_activity’

Die Ausgabe dieses Befehls findet ihr hier auf Pastebin.

Wie ihr sehen könnt, laufen zu diesen Zeiten etwa 32 UPDATE-Abfragen. Wenn das passiert, zeigt iotop an, dass die Datenbank zwischen 2,5 MB/s und 15 MB/s liest und schreibt.

Wenn ich die laufende UPDATE-Abfrage mit dem protokollierten Query (in /var/discourse/shared/standalone/log/var-log/postgresql) vergleiche, stelle ich fest, dass diese Abfragen extrem lange dauern:

2018-07-03 12:51:27.052 UTC [17504] discourse@discourse LOG:  duration: 2352061.872 ms  statement: UPDATE drafts
                       SET  data = '{"reply":"<redacted for debugging purposes>","action":"reply","categoryId":24,"postId":118034,"archetypeId":"regular","whisper":false,"metaData":null,"composerTime":65922,"typingTime":8400}',

entspricht

  discourse | 17504 | active | UPDATE drafts                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  +
           |       |        |                SET  data = '{"reply":"<redacted for debugging purposes>","action":"reply","categoryId":24,"postId":118034,"archetypeId":"regular","whisper":false,"metaData":null,"composerTime":65922,"typingTime":8400}',                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    +
           |       |        |                     sequence = 124,                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            +
           |       |        |                     revisions = revisions + 1                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                  +
           |       |        |                WHERE id = 84548

Wenn ich die Docker-App neu starte, kehren die Abfragen zwar zurück, aber im Grunde kann ich nur abwarten, was meine Nutzer unzufrieden macht.

Bitte gebt mir Bescheid, ob ich etwas tun kann, um das Problem zu mildern – zum Beispiel indem ich Wartungsaufgaben auf etwa 5 Uhr morgens verschiebe.

Wenn ihr weitere Vorschläge habt, wie ich tiefer graben kann, teilt sie bitte mit! Jede Hilfe ist willkommen

Ninja-Edit: Mir ist gerade noch eine zusätzliche Information eingefallen: Ich konnte dies nicht mit bestehenden Cronjobs in Verbindung bringen (weder vom Host-Betriebssystem noch innerhalb der Docker-App).

Einige Hintergrundinformationen zum Server:

• Server-Betriebssystem: Ubuntu 18.04LTS
• Es handelt sich um eine VM mit 100 GB Festplattenspeicher, 4 GB RAM und 4 Kernen
• Die Festplatte ist ein RAID10 mit 6 Festplatten und 15K Umdrehungen
• Discourse v2.1.0.beta2 +107
• Plugins: babble, whos-online, voting, cakeday, anonymous-categories, league
• Seitenaufrufe pro Tag: Ca. 75.000
• Nutzer: 1.350, davon täglich ca. 700 aktiv
• 73.600 Beiträge, 351 Themen, das aktivste Thema enthält 13.500 Beiträge

Danke, dass ihr das hier bis zum Ende gelesen habt!

[Falco]

@JanJoost

• Disk is a 6-disk RAID10 with 15K disks

• Are those spinning disks?

• How big is your database?

I had an experience like this with the combo: active forum + spinning disks. In the end it was Redis and PostgreSQL fighting for IO and after a while redis would stop because it had multiple rdb saves in the queue to happen.

Can you please check if redis is saving to disk during the problem?

[JanJoost]

They are SAS disks, 900GB 15K (spinning disks, not SSDs) configured in RAID10. They are not exactly slow, but you obviously wouldn’t get the IOPS you’d get from an SSD.

As for the database: the postgres_data directory is 4.5GB.

I doubt it’s redis in this case though - I stupidly enough forgot to grab a full screenshot from iotop but I can’t remember seeing redis in there. Did see postgres UPDATE at the top of the list with IO activity of 98% and up.

BTW: I’ll check tomorrow as I have no doubt it will happen again. And I’ll grab a screenshot from iotop.

[codinghorror (Jeff Atwood)]

@Falco

after a while redis would stop because it had multiple rdb saves in the queue to happen.

I still don’t feel redis should be constantly clobbering the disk. It’s supposed to be an in memory key-value store most of all.

[JanJoost]

@Falco

Can you please check if redis is saving to disk during the problem?

Hi, server’s very busy again - changed the timezone within the container from UTC to GMT+1 (Europe/Amsterdam) and indeed, load now starts 2 hours later.

Screenshot from iotop -o (which only lists the actual tasks doing I/O):

I’ve observed the behaviour and indeed, over the last 5 minutes, I did see Redis pop up twice, but only for a moment. Doesn’t come across as being the source of the problems to me.

I have a feeling there are some scheduled tasks which are being started, but I can’t really find the source. Checked cron jobs from both within the container and the underlying OS - no dice.

If you have any suggestions then those would be more than welcome! Also, if there’s more I can do regarding information gathering, please tell me!

[codinghorror (Jeff Atwood)]

More likely Postgres then, if you have ruled Redis out.

[JanJoost]

@codinghorror

More likely Postgres then, if you have ruled Redis out

I think so too - see the screenshot of the output of iostat - it’s a PostgreSQL UPDATE command that generates most of the IO load.

Thing is: what is triggered here? And why? Where can I find how (and when!) the postgresql maintenance tasks are triggered?

Edit: Additional info: Screenshot from sidekiq:

[Falco]

Looks like your problem is this one: Reschedule sidekiq scheduled job

You have one still active topic (last reply 19m ago) with 17k posts. That plus slow disks kill performance. The fact that your data is larger than the RAM also makes this more complicated.

[JanJoost]

@Falco

You have one still active topic (last reply 19m ago) with 17k posts.

That’s the most popular topic indeed. Would it be better to break it up in pieces, close this one and open a new one?

Thanks for the pointer btw - I’ll have a look!

[Falco]

@JanJoost

Would it be better to break it up in pieces, close this one and open a new one?

Yes. I set auto close topics post count to 1024 or 512 when I had an HDD instance.

Also, I completely forgot, but looks like I took an stab at this query 2 years ago:

Post.calculate_avg_time() taking up a long time

I believe the daily version of the query can be optimized by moving the filter up in the subquery. Before: kmV3 : Discourse Post.calculate_avg_time daily | explain.depesz.com After: i3JU : Discourse Post.calculate_avg_time daily Optimized | explain.depesz.com This would not apply to the weekly one, but as you stated the daily one is the bugger. Can you try that and confirm the differences @michaeld ?

[codinghorror (Jeff Atwood)]

Surprising that you are running into this with 17k posts in one topic, I view 20k as the “starting to be a problem” cutoff. But maybe we should set the bar lower cc @tgxworld

[JanJoost]

@codinghorror

Surprising that you are running into this with 17k posts in one topic

It could be related to my specific setup though - if it helps i can document a bit more, run some additional tests etc (bonnie++ or so).

[Falco]

@JanJoost

if it helps i can document a bit more, run some additional tests etc (bonnie++ or so).

We have some public results about disk bench here:

@codinghorror

Disk benchmarks

ioping -RD -w 10 .

12.8k iops, 50.0 MiB/sec vs 13.6k iops, 53.2 MiB/sec

(roughly the same disk performance)

Maybe run those (when the forum has low load) to compare?

[RGJ (Richard - Communiteq)]

Just as a thought exercise: what exactly would break if this specific job would not exist or run?

I can’t find any place in the code where Post.avg_time and Topic.avg_time are actually being used? Or is there maybe some generated accessor that I’m missing?

[zogstrip]

It’s used to compute the post score that we use to rank posts so we can summarize them.

[sam (Sam Saffron)]

@codinghorror

Surprising that you are running into this with 17k posts in one topic, I view 20k as the “starting to be a problem” cutoff. But maybe we should set the bar lower cc

I have mentioned this particular issue before… it is totally unrelated to @tgxworld’s work.

Reschedule sidekiq scheduled job

If we give up the requirement for a gemetric mean of read time on every post on large topics, we can make the job much faster and then its a lot less likely to kill his system at 3AM. In theory that code should work, my guess is that it is writing it to the wrong part of redis, you want to be at the default site of the multisite.

I am still not buying that we are getting any value from “geometric mean of all read times for all posts”… a simple… which posts are read most is probably good enough for the “best of” algorithm.

Plus this whole concept is a bit broken imo cause the further you get to the end of a topic the less people read it.

I think a simple fix here, is to stop doing the whole geometric mean read time thing on topics with more than say 1000 posts. Which is oddly the only time we would use it…

[JanJoost]

@sam

I think a simple fix here, is to stop doing the whole geometric mean read time thing on topics with more than say 1000 posts. Which is oddly the only time we would use it…

Silly idea perhaps: Topics have a certain flow, setup, experience to them. Some topics are by nature filled with short posts, others with long posts.

Would it be an idea to calculate the mean over the first 500 or 1000 posts and use that as an indication for all posts in the thread?

After all, it’s an indication so an approximation would be acceptable, no?

[RGJ (Richard - Communiteq)]

@JanJoost

calculate the mean over the first 500 or 1000 posts

I don’t even know if this would help in our case. Fact is that post_timings is bigger than the internal memory of the server (20 GB vs 16 GB) and even limiting to the first x posts would still result in the database failing to pull the entire table into memory.

So my feeling is that attempting to do this query differently would work better than trying to “limit” it.

[JanJoost]

@Falco

Disk benchmarks

ioping -RD -w 10 .

12.8k iops, 50.0 MiB/sec vs 13.6k iops, 53.2 MiB/sec

(roughly the same disk performance)

Maybe run those (when the forum has low load) to compare?

Hi, I did run a test just now on the server (still relatively active with 30 active users online). These are the results:

19.3 k requests completed in 9.89 s, 75.5 MiB read, 1.95 k iops, 7.63 MiB/s
generated 19.3 k requests in 10.0 s, 75.5 MiB, 1.93 k iops, 7.55 MiB/s

Everything seems to be about 1/7th of the speed you get. Perhaps indeed I should consider switching from old-school HDD to SSDs.

[codinghorror (Jeff Atwood)]

These numbers are terrifyingly bad even for traditional HDDs. Way below what fancy enterprise 10k and 15k drives should be doing.